An algorithm for characterizing context-governed speech production patterns

Abstract

Speech recognition and analysis can be improved by using methods that can effectively characterize important speech patterns of a speaker without requiring hours of data. This thesis defines a method by which key contexts related to systematic speech modification can be used to create a profile of the speech produced by a speaker. Using acoustic and prosodic information, contexts that create the potential for speech modifications can be specified. Then, by filtering speech produced by a speaker in the targeted contexts, the patterns of speech production in these contexts can be characterized. With these productions, likely underlying contexts that are associated with the productions can be used to enhance speech recognition when these contexts arise in new speech.